Evidence before enthusiasm

Know what every claim means.

Useful decisions need visible sources, assumptions and limitations.

Specifications are not performance.

Manufacturer specification

A published product configuration, linked to the original source. Exact SKUs, regional options and memory differences matter.

Contributor / founder-reported

An attributed observation supplied by its author. Publishing does not establish independent verification.

Lab-verified · future workflow

Reserved for a documented reproduction with pinned setup, raw evidence and review. No performance result published so far carries this label.

Planning scenario

Editable assumptions used to explore a decision. This includes the cost calculator, software allowances and finance scenarios. Published catalog offers are separate, dated supplier snapshots.

MEMBER RATINGS

Ratings are opinions, labelled as such.

Signed-in members can rate each machine once, from one to five stars, and say whether they run it, evaluated it hands-on or are exploring. That relationship is self-declared and shown with the rating. Stars enter the average immediately; written verdicts are published after a review by the Lab team, which checks for confidential information, supplier self-promotion and abuse, not for accuracy of opinion. A rating never stands in for a measured run.

The first-pass screen

Make the memory assumption explicit.

weights = parameters × effective bits ÷ 8 × 1.03 · KV = KB/token × context × requests · reserve = 6 GB (discrete) or 10 GB (unified) · fits if total ≤ 90% of the memory tier

Effective bits include quantisation scales (4-bit formats count 4.6 bits). KV bytes per token come from each model card: 2 × layers × KV heads × head size × 2 bytes, halved for an 8-bit KV cache. The speed ceiling is bandwidth ÷ bytes read per generated token (active weights plus one request’s KV cache); batched streams share the weight read. These are first-order estimates: runtimes, kernels, prefill and scheduling change the real numbers, so every screen ends in a measured run.

The screen compares this estimate with 75% of listed memory. For GB300 it uses only the 252 GB GPU figure; CPU offloading needs separate validation. HP’s system memory figure is not a promise that the same amount is available to its GPU.

Eligibility also checks requested CUDA. A EUR budget is checked only against a published EUR offer; other currencies and quote-only machines require a separate price check. Results prioritize fewer failed technical constraints, with catalog order as a tie-breaker. They do not rank measured speed.

Try the finder

Who publishes the Lab.

OnPremBench is published by Understand Tech. Understand Tech also sells AI-in-a-Box, a managed deployment on NVIDIA GB10 and GB300 appliances, described on its own page. That offer does not change the records: specifications, prices, evidence labels and planning figures follow the same rules for every manufacturer, the catalogue opens sorted by price, and no machine is placed or ranked for payment. Understand Tech’s own deployments appear as use cases labelled founder-reported, like any other attributed source.

Supplier offers and planning scenarios are separate.

Catalog prices are snapshots taken from linked supplier pages on 7 September 2026. Each offer states its currency, region, selected configuration and any published tax treatment. They are not live inventory feeds or UT offers. A missing, zero or ambiguous store price is shown as a quote request. DIY kits and integrated workstations are identified separately. Software, support and setup allowances remain editable examples, not UT pricing.

The calculator uses an independent, editable scenario in the currency you choose (USD by default). It does not convert foreign prices or use PSU ratings as average power. Enter your actual quote and measured consumption. It includes equipment, setup, electricity, software/support and administration, less end-of-horizon residual value. Compare with a cloud option meeting the same quality, demand and availability. Hardware you already own can start at zero incremental purchase cost; include depreciation or replacement costs if relevant to your decision.

The finance scenario uses a 6% nominal annual interest rate with no fees, deposit or residual. It is not a lender offer. Eligibility and commercial terms remain unconfirmed.

Local inference may avoid a per-token API charge, but the system still has equipment, electricity and operating costs. Token speed is different from financial token consumption.

Reviewed 7 September 2026

Real systems. Traceable sources.

NVIDIA DGX Spark

Memory capacity does not establish interactive speed or simultaneous request capacity. UT has documented a Spark deployment. Confirm exact software versions and support scope.